Multimodel Ensembles of Streamflow Forecasts: Role of Predictor State in Developing Optimal Combinations

Seasonal streamflow forecasts based on climate information are essential for short-term planning and for setting up contingency measures during years of extreme climatic conditions. Recent research shows that operational climate forecasts obtained by combining different General Circulation Models (GCM) have improved predictability/skill in comparison to the predictability that could be obtained from single GCM. In this study, we present a new approach for developing multi-model forecasts that combines streamflow forecasts from various models by evaluating their skill from the predictor state space. Based on this, we show that any systematic errors in model prediction with reference to a particular predictor conditions could be reduced by combining forecasts from multiple models along with climatological ensembles. The methodology is demonstrated through development of multi-model ensembles of streamflow forecasts for the Falls Lake reservoir in Neuse river basin, NC by combining probabilistic streamflow forecasts from two low dimensional statistical models that uses SST conditions in Tropical Pacific, North Atlantic and North Carolina Coast as predictors. Using Rank Probability Score (RPS) for evaluating the predictability of seasonal (July- August-September) streamflow forecasts available each year from the two candidate low dimensional models, the methodology proportionately gives higher representation by drawing increased ensembles for a model that has better predictability under similar predictor conditions. The performance of the multi-model forecasts are compared with the individual model's performance using various performance evaluation measures such as correlation coefficient, root mean square error (RMSE), average Rank Probability Skill Score, average Rank Probability Skill Score (RPSS) and reliability diagrams. By developing multi-model ensembles for leave-one out cross validated forecasts and adaptive forecasts based on the proposed methodology, the study shows that evaluating the model's performance based on the predictor state provides a better alternative in developing multi-model ensembles instead of combining models purely based on their long-term predictability.